Our bias review happens after the model ships. Where does yours sit?

3 weeks after our last launch, our routine metrics review flagged that new sellers were recommended 23% less often than established sellers with comparable ratings.

The ranking change itself was very small. We had tuned for click-through and it worked. But nobody checked subgroup selection rates pre-launch because our fairness review sits with a separate governance group. We only hear from them post-launch.

We rolled back that ranking weight and patched the monitoring. Still, it bothers me that detection depends on us catching it in post-hoc metrics.

Where does fairness review sit in your org, inside the build team or outside it? Has it ever actually blocked or meaningfully changed a launch?

separate gate with veto wins. embedded reviewers fold under launch pressure. fraud taught me.

embedded reviewers fold under launch pressure

Our separate gate has veto on paper and has never used it. Last quarter it approved our model despite a gap for new businesses because targets were on the line. The only change came from an analyst in sprint review who flagged the cutoff early. Who pays the cost when the gate says no the week of launch?

Independent validation.

Who pays the cost when the gate says no the week of launch?

In banking the business line pays in schedule. Validators sit outside the build org and nothing goes live without sign-off under SR 11-7, Fed and OCC 2011 model-risk guidance.

Now carried in April 2026 interagency guidance, OCC Bulletin 2026-13. This writeup nails it: “The review gate should block the “ready to ship” status the same way a failing test blocks a merge.” A veto that never fires is staffing, not structure.